THE HITL PARADIGM, FROM A TO Z
A few weeks ago we talked about the revolution being undergone by the implementation of Artificial Intelligence solutions, accelerated by digital transformation as the only way to sustain business in the times of Covid-19 and speed the recovery of economies post-pandemic, as per accurate data from IDC.
We offered the notion of HITL (Human in the Loop) as one of the paradigms involved in incorporating human feedback during the training (modeling and simulation) stage of machine learning algorithms.
NOW – WHAT ARE HITL’S PURPOSE?
This approach is gaining ground as the best possible way to train more accurate models, insofar as people continually test, adjust and feed data, in a live, constructive and virtual manner.
Since machine-driven systems haven’t yet reached the desired accuracy levels, human intervention in training circuits is required in order to create more accurate automatic learning levels. In other words, HITL describes the procedure carried out in cases when the machine or computer system can’t offer an answer to a problem and requires human intervention.
REDUCING COSTS AND NEW DEVELOPMENTS
Automatic learning models are essential to any company intending to develop an innovative business, since they enable the automation of different routine aspects of procedures and circuits. But actually they are still expensive to build, train, validate and curate, since in order to manage the process data scientists are required – and their salaries are not exactly affordable. This becomes even more complicated when data are not structured, as happens in 90% of cases. That’s why it’s so important to optimize these procedures – and that’s why we at Arbusta develop technological services that revolve around this paradigm, and seek to report everything about HITL for companies.
HITL is a paradigm involving IT talent, since it combines supervised machine learning (which works with labeled or selected data sets to train algorithms through parameter adjustment) and active learning (which feeds data back into a classifier). The aim of this HITL paradigm is to correct inaccuracies in the machine’s predictions – and this requires human intervention.
Many data scientists employ the approach known as the “80/19/1” equation, which establishes that, in 80% of the cases, the algorithm must be left alone to learn, while humans must intervene in 19% of the cases and the remaining 1% must be left to chance. With the HITL approach, on the other hand, people intervene when the machine is not sure of the answer. In these cases, human criterion is resorted to and then added to the model as feedback for optimization. But, in specific cases, when is the human perspective resorted to?
- When the algorithms don’t understand the information from the start, or don’t know how to perform the task.
- When the cost of error is too high (in industries such as insurance or the health sector, for example).
- When what one is looking for is quite unusual or rare.
- When there are still few data available on a given subject or business.
- When the entry data is incorrectly interpreted.
THE QUEST FOR ACCURACY
As a consequence of these interventions, the HITL approach reduces training times and enables the obtainment of more accurate data. In which specific cases? When working with chatbots in first-level customer service instances, for example. In such circumstances, customers sometimes give too many details and human intervention is needed. Achieving progress in this is key, especially when we consider that 25% of customer service and support transactions will be part of the virtual assistant or chatbot technology by 2020, as compared to less than 2% in 2017.
The HITL paradigm also applies to the training of autonomous vehicles: in this area, algorithms today score right driving answers in 90% of cases, a percentage that is certainly high but not sufficient, since there are human lives at stake. HITL can be used in other cases as well, for example for optimizing the algorithms of security cameras reacting to a movement sensor (to reduce false alarms), or in text messaging apps that transcribe voice to text with high precision (to solve problems posed by a specific jargon).
In the case of chats and virtual assistants, Arbusta’s IT talent is focused on training the answers natural to certain dialogues, entities and intentions. We also allocate metadata to images or to digital video, through keywords or subtitles, and facilitate training and adjustment in text recognition contexts.
The HITL paradigm is destined to make automatic learning apps cheaper and more accessible – and, in that regard, the IT talent of consultancy companies such as Arbusta has a lot of value to add.